Keynote: "Cloud-of-Clouds Storage: from Theory to Production" (joint with DLT4SEC)
Alysson Bessani (FCUL-LASIGE)
Bio
Alysson Bessani is a Professor at the University of Lisbon Faculty of Sciences, Portugal, and the director of the LASIGE research unit. He received his Ph.D. in Electrical Engineering from UFSC (Brazil) in 2006, was a visiting professor at Carnegie Mellon University (2010), and a visiting researcher at Microsoft Research Cambridge (2014). Alysson coordinated/collaborated on 13 international projects and co-authored more than 120 peer-reviewed publications on dependability, security, Byzantine fault tolerance, and cloud storage. His work has been recognized with multiple awards, including two DSN test-of-time awards. He is also the original developer of the BFT-SMaRt replication library (http://bft-smart.github.io/library/) and one of the co-founders of the Vawlt cloud storage startup (https://vawlt.io). More information about him can be found at http://www.di.fc.ul.pt/~bessani.
Abstract
Public clouds have become a fundamental pillar for many companies and internet-scale services. In particular, the exponential rise in the amount of generated data made (public) cloud storage services a key infrastructure for dealing with such data. Despite the huge efforts of cloud providers to secure their services, there is a profusion of events related to intrusions, unavailability, and even data corruption/destruction. One way to minimize the effect of such events is to exploit the existence of a vibrant cloud market and store data on a set of different providers, implementing a cloud-of-clouds storage strategy. To do that, cloud users must rely on fault-tolerant storage abstractions built on top of fail-prone individual cloud services, namely read/write registers and lease objects. Many proposals for such abstractions come from both the theoretical and the systems communities. Based on a subset of these works, we select and present four constructions (three types of registers and one lease object), comparing their properties and costs, and discuss how they were used to implement practical cloud-of-clouds storage services (from a programming library to a commercial storage service, vawlt.io). In the end, we’ll also share some of our experience in transforming fundamental research done in academia into a startup that secured more than 3M euros of VC funding.
13:00 — 14:00
Lunch Break
14:00 — 14:10
Welcome and Introduction
14:10 — 15:10
Session 1: Optimizations for storage
EDIC: Towards Improving Fault Tolerance of Erasure Codes with Redundancy Conversion Ioannis Tzouros and Vana Kalogeraki (AUEB)
Abstract
To adapt to the growth of data access resiliency, distributed cloud storage systems are increasingly deploying erasure coding methods with redundancy conversion which adjust their parameters to encode and repair constantly expanding data sets. However, while new parity blocks are generated for updated data groups, any parity blocks created from previous data group updates are not exploited further, or even entirely discarded. This paper proposes EDIC (Elastic Diagonally Interleaved Coding), a middleware that optimizes fault tolerance for converted data groups. EDIC deploys diagonally interleaved coding, an advanced erasure coding method, enhanced by leveraging multiple versions of parity blocks, inspired by the principles of Elastic Reed-Solomon codes. EDIC exploits the parity blocks of a previous version of a converted data group to repair the original, non converted blocks within the group, while the new parity blocks will cover the recently added blocks, incurring only a minimal storage overhead.
Speed Kills: Revisiting Data Deduplication for Modern Storage Devices Rui Pedro Oliveira, Tânia Esteves and João Paulo (INESC TEC & U. Minho)
Abstract
The rise of new, faster storage devices has introduced significant changes to the design and optimization of traditional storage systems. This paper examines how different storage hardware, ranging from HDDs to SSDs and Persistent Memory, influences the design and efficiency of modern deduplication systems. Through a literature consolidation and an empirical study on the performance implications of different fingerprinting methods and indexing configurations, we highlight important and non-trivial co-design choices that must be carefully considered when building next-generation deduplication systems.
Abstract
Programmable caching engines like CacheLib are widely used in production systems to support diverse workloads in multi-tenant environments. Its design focuses on performance portability and configurability, allowing applications to inherit caching improvements with minimal implementation effort. However, its behavior under dynamic and volatile workloads remains largely unexplored. This paper presents an empirical study of CacheLib with multi-tenant settings under dynamic and volatile environments. Our evaluation across multiple CacheLib configurations reveals several limitations that hinder its effectiveness under such environments, including rigid configurations, limited runtime adaptability, lack of quality-of-service support and coordination, which lead to suboptimal performance, inefficient memory usage, and tenant starvation. Based on these findings, we outline future research directions to improve the adaptability, fairness, and programmability of future caching engines.
15:10 — 15:15
Short Break
15:15 — 16:05
Session 2: Storage for Applications
PolyLayer: the next 700 storage configurations João Lopes (InvisibleLab), Bruno Pereira (INESC TEC & U. Minho), Filipe Pereira (INESC TEC & U. Minho), Vicente Muñoz (InvisibleLab), Tiago Gomes (InvisibleLab), Rui Ribeiro (InvisibleLab), Filipe Costa (InvisibleLab), Marta Bonjardim (InvisibleLab), Francisco Cruz (InvisibleLab), João Paulo (INESC TEC & U. Minho) and Francisco Maia (U. Porto & INESC TEC)
Abstract
Modern storage systems requirements demand flexible, scalable solutions that address diverse concerns such as data reduction, replication, security, and multi-cloud distribution. Existing solutions often provide these guarantees through monolithic implementations, limiting their adaptability to specific application needs. This paper introduces PolyLayer, a multi-interface, composable and multi-backend storage architecture. It builds on the concept of stackable storage architectures and redesigns these to support commonly used user APIs (e.g., POSIX, Key-value, Object store), while providing support for data persistence across multiple storage backends (i.e., on-premises, cloud services, blockchain). We present the first steps towards the design of such architecture, while implementing a proof-of-concept and evaluating it. Our preliminary results show that the design can effectively be used in real-world scenarios where new functionality is added to a storage system with low overhead over the base system. For instance, we show how anti-tampering mechanisms can be added to a traditional relational database without any change to the database itself or the application using it.
Abstract
This paper presents the first study on distributed cache solutions for Distributed Deep Learning training in HPC centers. We categorize these according to their intrusiveness, cache design and resources, sample placement and retrieval. Further, through an empirical evaluation on real supercomputers (Deucalion and Vista), we measure the impact of caching and checkpointing data into different storage sources (i.e., parallel file system, compute node’s local storage, remote nodes). Results show that choosing the optimal caching design is dependent on the targeted infrastructure, the storage resources used for caching data, and the network protocols. Based on these, we outline design trade-offs and propose future directions for co-designing cache systems within HPC infrastructures to more effectively support Distributed Deep Learning workloads.
PADME: Probabilistic Data Management for Efficient ML/AI (Fast Abstract) Francisco Maia and Carlos Baquero (U. Porto & INESC TEC)
Abstract
The exponential growth of data and the increasing complexity of machine learning (ML) and artificial intelligence (AI) workloads have exposed the limitations of current large-scale data management systems. This paper introduces PADME, a research initiative to develop a probabilistic data management system tailored for efficient, scalable, and resilient ML/AI applications. PADME leverages unstructured epidemic protocols and adaptive system architectures to address the challenges of data persistence, fault tolerance, and dynamic workload adaptation in distributed environments. We outline the motivation, related work and current approaches, and the high-level design for PADME.
16:05 — 16:10
Closing Remarks
Note: The presentation time for research papers is 15 minutes, followed by a 5-minute discussion. Fast abstracts have 7 minutes for presentation and 3 minutes for discussion.
